Neutralization avoidance and naturalness in the learning of palatalization

Heng Yin James White

University College London

Learning bias

u  Learners exhibit a variety of learning biases when learning phonological alternations. (see Moreton & Pater, 2012a/b for a summary)

u Most previous studies looking at learning biases focused on non-

neutralizing alternations. (e.g., Wilson, 2006; Peperkamp & Dupoux, 2007; White, 2014)

u  Neutralizing alternations?

2

Neutralization avoidance (contrast preservation)

u  Previous studies have appealed to neutralization avoidance in analyses of phonological patterns:

l  Diachronic perspective: selection process against neutralizing patterns, especially those resulting in ambiguous speech; functional load hypothesis (e.g., Wedel, 2006; Silverman, 2010; Wedel et al., 2013)

l  Synchronic perspective: •  MAXIMIZE CONTRASTS (Flemming, 1996, 2004) •  *NEUT (Bolognesi, 1998) •  *MERGE (Padgett, 2003, 2009) •  PRESERVECONTRAST (Lubowicz 2007)

3

Some neutralization avoidance effects l  Preservation of contrasts with phonetic shifts. (Flemming, 1996)

l  English: Difficult to maintain voicing in initial stops, so /b/ à [p], /p/ à [ph].

l  Rules applying to an isolated sound rather than a natural class of sounds. l  E.g. stop nasalisation in traditional Tokyo dialect of Japanese: (Labrune 2012)

[g] → [ŋ] *[d] → [n], *[b] → [m] /n/ and /m/ are phonemes in Japanese while [ŋ] is not.

l  Exceptional allomorph selection just in case it would avoid neutralization. l  Polish (Lubowicz, 2007)

l  Optional rules less likely to apply if they might neutralize. l  Japanese nasal contraction (Kaplan & Muratani, 2015)

4

Research question u  A lot of work discussing the role of the speaker and listener in

neutralization avoidance. (see Silverman 2012 for an overview) u But not much work looking at the role of the learner.

u Question: Are learners biased against neutralizing alternations in

comparison with non-neutralizing ones?

u  Artificial language approach: u Present equal input for neutralizing and non-neutralizing

alternations. u Test how well learners acquire the two types of alternation.

5

Experiment 1: method u  Participants: native English speakers (n=30)

u  3 phases:

l  Exposure phase l  Test phase 1: trained items l  Test phase 2: novel items

u  4 novel alternations involving palatalisation [t, d, s, z] ~ [tʃ, dʒ, ʃ, ʒ]

u  2 counterbalancing groups: Language A vs. Language B 6

7

Language A

Experiment 1: design Language B

Alternations

/ t / ⟶ [ tʃ ]

Critical non-alternating phonemes / tʃ, dʒ / / ʃ, ʒ /

Filler non-alternating phonemes / p, b, k, g, f, v/ / p, b, k, g, f, v/

/ d / ⟶ [ dʒ ]

/ s / ⟶ [ ʃ ]

/ z / ⟶ [ ʒ ]

/ t / ⟶ [ tʃ ]

/ d / ⟶ [ dʒ ]

/ s / ⟶ [ ʃ ]

/ z / ⟶ [ ʒ ]

Experiment 1: stimuli u  Exposure stimuli: 48 CVCVC singular nonwords with CVCVC-i plural

forms: l  8 alternating [t ~ tʃ] and [d ~ dʒ] (Neutralizing in Language A)

[tusut] ⟶ [tusutʃi] l  8 alternating [s ~ ʃ] and [z ~ ʒ] (Neutralizing in Language B)

[duvis] ⟶ [duviʃi]

l  8 critical non-alternating trials ending in [tʃ, dʒ] (Language A) or [ʃ, ʒ] (Language B) [buvatʃ] ⟶ [buvatʃi]

l  24 non-alternating filler trials ending in [p, b, k, ɡ, f, v]. [vatuk] ⟶ [vatuki]

8

Experiment 1: stimuli u  ‘Ilegal’ sequences never presented.

u  *[ti, di] in Language A. u  *[si, zi] in Language B.

u Otherwise, consonant and vowel distribution roughly balanced across positions.

9

Experiment 1: exposure phase

10

Experiment 1: test phases u  2 test phases: 24 trained items, then 48 untrained items.

u  Forced-choice task: choose the correct plural form between an alternating option and a non-alternating option.

11

Experiment 1: test phase options u  Incorrect changing options for non-alternating phonemes:

12

/ tʃ / ⟶ [ ʃ ]

/ dʒ / ⟶ [ ʒ ]

/ ʃ / ⟶ [ tʃ ]

/ ʒ / ⟶ [ dʒ ]

/ p / ⟶ [ tʃ ]

/ b / ⟶ [ dʒ ]

/ k / ⟶ [ tʃ ]

/ g / ⟶ [ dʒ ]

/ f / ⟶ [ ʃ ]

/ v / ⟶ [ ʒ ]

40%

50%

60%

70%

80%

90%

Neutrailizing Non-neutralizing

Mea

n ac

cura

cy (%

)

Experiment 1: results (Neut. vs. Non-neut. overall)

Better learning in Non-neutralizing condition.

*

13

40

50

60

70

80

90

Trained Novel

Mea

n ac

cura

cy (%

)

Neutralizing,

Non.neutralizing,

Experiment 1: results (trained and new items)

* * * marg. Main effect of training.

Trained > New

Main effect of Trial Type. Non-neut. > neut.

No interaction.

14

Summary 1.  Neutralizing alternations dispreferred relative to Non-neutralizing

alternations, despite equal evidence for both. •  Independent of which alternations were being learned.

15

0%

10%

20%

30%

40%

50%

60%

Velar Labial

Mea

n ac

cura

cy (%

) A second interesting result: velar vs. labial fillers

More errors on velar fillers than labial fillers!

*

16

Summary 1.  Neutralizing alternations dispreferred relative to Non-neutralizing

alternations, despite equal evidence for both. •  Independent of which alternations were being learned.

2.  Participants spontaneously palatalized velar stops more often than labial stops.

•  Even though neither was palatalized in exposure.

17

Why velars more than labials?

u  Substantive naturalness bias? u Typologically, palatalization of velars more common than labials. (Kochetov

2011) u  [ki] and [tʃi] are more phonetically similar than [pi] and [tʃi], and thus

considered a more likely alternation by learners. (e.g. Steriade, 2001) u Consistent with previous studies (Wilson, 2006; Skoruppa et al., 2011; White, 2014;

White & Sundara, 2014)

u  L1 influence? u English has alternations involving velar palatalization (‘velar softening’).

u opaque [oʊˈpeɪk], opacity [oʊˈpæsɪɾi]; esophagus [əˈsɑfəgəs], esophageal [ə̩sɑfəˈdʒiəl] (Halle 2005)

u Cross-linguistic study would be useful. 18

Accounting for the anti-neutralization effect u Why are neutralizing rules more difficult to learn?

u We consider two possibilities: •  pure statistical learning •  learning bias

19

Just distributional learning?

Non-neutralizing alternation

Neutralizing alternation

Non-alternating fillers

[s] [ʃ]

_i _a _u

[t] [tʃ]

_i _a _u

[p] [tʃ]

_i _a _u

Fully complementary distribution

Partially overlapping distribution

Fully overlapping distribution

20

Discussion: distributional learning u  A distributional learning model must be able to recognise both fully

complementary distributions and partially overlapping distributions as cues for alternation. (Calamaro & Jarosz, 2015)

u  But perhaps identifying partially overlapping distribution requires more input or computation, resulting in reduced learnability.

u  Potential limitation: Is there more than just phonological distributions involved? Lexical, semantic, and/or pragmatic influences? u  If so, this is unlikely to be the full story.

21

Anti-neutralization learning bias u  Learners have a neutralization avoidance bias when learning alternations.

u  Could be formalized in a learning model with weighted constraints (e.g. MaxEnt). u A prior could initially assign the ANTI-NEUTRALIZATION constraint a high weight.

u  Would block alternations resulting in neutralization. u Exposure to neutralizing alternations would gradually lower the constraint’s

weight. u Would make neutralizing alternations learnable, but at a slower rate.

22

What type of constraint/bias? •  Phonological

–  Does it consider all possible forms (e.g., Lubowicz, 2007; Padgett, 2009) or only existing forms in the lexicon?

–  Is it restricted to paradigms, or can it consider any pair of forms? (see Kaplan & Muratani, 2015 for discussion)

•  General/functional –  How do we integrate the neutralization avoidance pressure with models of the

grammar (assuming we want to do that)? (e.g. Flemming 1996, 2004) –  What roles does homophony (actual or potential) or frequency play? (Wedel et al.

2013)

•  Upshot: If we see neutralization avoidance effects in learning, learning experiments could provide us with a new way of looking at these questions.

23

Is neutralization avoidance driven by homophony avoidance?

u  Phonologically neutralizing alternations tolerated when they result in little lexical neutralization. (e.g. Silverman 2010 on Korean)

u  Diachronically, mergers less likely when they would result in high amounts of homophony. (Wedel et al. 2013)

u  Synchronically, stochastic processes may occur less frequently when they result in potential homophones. (e.g. Kaplan & Muratani, 2015 on Japanese nasal contraction)

24

Does homophony affect the learnability of neutralizing alternations? •  Currently running Experiment 2 to address this question.

Exp. 1: Half lexical neutralization

Exp. 2: Homophony Condition Exp. 2: No Homophony Condition

25

Product-oriented learning? •  Product-oriented learning (e.g. Bybee & Slobin, 1982) and neutralization

avoidance predict different results in our experiment: –  Product-oriented learning: Adding /tʃ/ ⟶ [tʃ] should boost /t/ ⟶ [tʃ] due

to more cases of plural -tʃi. –  Neutralization avoidance: Adding /tʃ/ ⟶ [tʃ] should reduce /t/ ⟶ [tʃ].

•  We found neutralization avoidance, not product-oriented learning. –  Fillers further suggest that product-oriented generalization was not a major factor

in our study.

•  Results diverge from Kapatsinski 2013. –  He found that adding /tʃ/ ⟶ [tʃ] had no impact on (trained) /k/ ⟶ [tʃ]. (no

neutralization avoidance) –  But it did increase (untrained) /t/ ⟶ [tʃ]. (product-oriented learning)

26

Conclusions u  Learners are biased against neutralizing alternations.

u Use of neutralization avoidance in synchronic accounts is potentially warranted.

u Learner could play a role in diachronic neutralization avoidance effects. u  Learners favor palatalization of velar stops over that of labial stops.

u Suggests a substantive bias, though a L1 bias cannot be ruled out at in this experiment.

27

Future directions u  Running the study with different L1 backgrounds.

u Would we see the same effect with infants? u  Perhaps related to the Mutual Exclusivity Bias in pragmatic development?

(Merriman, Bowman, & MacWhinney, 1989) l  Infants disfavour the learning of homophones. l  Neutralization avoidance could be a realization of mutual exclusivity at a

phonological level. l  Infants have ‘proto-lexical’ knowledge (Ngon et al., 2013) and phonotactic

generalizations (Mattys et al.1999) at an early age. u  L2 acquisition of natural languages.

l  L2 learners care more about semantic information than ALL participants. l  More explicit training on phonological system.

28

Acknowledgements: u  Help with experiment: Andrew Clark.

u  Helpful discussion: Wing Yee Chow, Richard Breheny.

u  London Phonology Seminar audience.

29

References: Bolognesi, R. (1998). The phonology of Campidanian Sardinian: A unitary account of a self-organizing structure. NOR. Bybee, J. L., & Slobin, D. I. (1982). Rules and schemas in the development and use of the English past tense.

Language, 265-289. Calamaro, S., & Jarosz, G. (2015). Learning General Phonological Rules From Distributional Information: A

Computational Model. Cognitive science, 39(3), 647-666. Flemming, E. (1996). Evidence for constraints on contrast: The dispersion theory of contrast. UCLA Working Papers in

Phonology, 1 (pp. 86-106).Flemming, E. (2004). Contrast and perceptual distinctiveness. In B. Hayes, R. Kirchner, and D. Steriade (Eds.)

Phonetically-Based Phonology (pp. 232-276). Halle, M. (2005). Palatalization/velar softening: What it is and what it tells us about the nature of language. Linguistic

Inquiry, 36(1), 23-41. Kapatsinski, V. (2013). Conspiring to mean: Experimental and computational evidence for a usage-based harmonic

approach to morphophonology. Language, 89(1), 110-148. Kaplan, A., & Muratani, Y. (2015). Categorical and gradient homophony avoidance: Evidence from Japanese. Laboratory

Phonology, 6(2), 167-195. Kochetov, A. (2011). Palatalisation. In M. Oostendrop, C. Ewen, B. Hume, & K. Rice (Eds.). Companion to Phonology.

(pp.1666-1690). Oxford: Wiley Blackwell. Labrune, L. (2012). The phonology of Japanese. Oxford: Oxford University Press. Łubowicz, A. (2007). Paradigmatic contrast in Polish. Journal of Slavic linguistics, 15(2). Mattys, S. L., Jusczyk, P. W., Luce, P. A., & Morgan, J. L. (1999). Phonotactic and prosodic effects on word

segmentation in infants. Cognitive psychology, 38(4), 465-494.

30

References: Merriman, W., Bowman, L., & MacWhinney, B. (1989). The Mutual Exclusivity Bias in Children's Word Learning.

Monographs Of The Society For Research In Child Development, 54(3/4), i. Moreton, E., & Pater, J. (2012a). Structure and Substance in Artificial-phonology Learning, Part I: Structure. Language

And Linguistics Compass, 6(11), 686-701.Moreton, E., & Pater, J. (2012b). Structure and Substance in Artificial-Phonology Learning, Part II: Substance. Language

And Linguistics Compass, 6(11), 702-718. Ngon, C., Martin, A., Dupoux, E., Cabrol, D., Dutat, M., & Peperkamp, S. (2013). (Non) words,(non) words,(non) words:

evidence for a protolexicon during the first year of life. Developmental Science, 16(1), 24-34. Padgett, J. (2003). Contrast and post-velar fronting in Russian. Natural Language & Linguistic Theory, 21(1), 39-87. Padgett, J. (2009). Systemic contrast and Catalan rhotics. The Linguistic Review, 26(4), 431-463. Peperkamp, S. & Dupoux, E. (2007). Learning the mapping from surface to underlying representations in an artificial

language. In J. Cole & J. Hualde (Eds.). Laboratory phonology (Vol. 9, pp. 315-338). Berlin: Mouton de Gruyter. Silverman, D. (2010). Neutralization and anti-homophony in Korean. J. Ling., 46(02), 453-482. Silverman, D. (2012). Neutralization. Cambridge: Cambridge University Press. Skoruppa, K., & Peperkamp, S. (2011). Adaptation to novel accents: Feature-based learning of context-sensitive

phonological regularities. Cognitive Science. 35, 348-366. Steriade, D. (2001). Directional asymmetries in place assimilation: a perceptual account. The role of speech perception

in phonology, 219-250. Wedel, A. (2006). Exemplar models, evolution and language change. The Linguistic Review, 23(3).

31

References: Wedel, A., Jackson, S., & Kaplan, A. (2013). Functional load and the lexicon: Evidence that syntactic category and

frequency relationships in minimal lemma pairs predict the loss of phoneme contrasts in language change. Language and speech, 0023830913489096.

White, J. (2014). Evidence for a learning bias against saltatory phonological alternations. Cognition, 130(1), 96-115. White, J., & Sundara, M. (2014). Biased generalization of newly learned phonological alternations by 12-month-old

infants. Cognition, 133(1), 85-90. Wilson, C. (2006). Learning phonology with substantive bias: An experimental and computational study of velar

palatalization. Cognitive Science, 30(5), 945-982.

32